AI Council: multi-agent debate with a live ChatGPT interface

The problem

A single AI response gives a user little visibility into alternative approaches or the assumptions behind the recommendation. For open-ended questions, seeing proposals challenged and revised can be more useful than receiving an immediate conclusion.

AI Council makes that process explicit. A Strategist, Pragmatist, Critic, Innovator and Philosopher answer the same question, review one another's proposals and vote after refining their positions.

My contribution

I built AI Council as an open-source MCP project with a live interface inside ChatGPT. The work connects a multi-provider agent layer, a structured debate coordinator and a React widget that makes the intermediate results visible.

The MCP tool starts the debate. Server-side state tracks messages and votes, while the widget receives updates so a user can follow the discussion and inspect the final result.

Engineering decisions

The debate has four explicit stages: proposal, critique, refinement and vote. Agents in the same stage run concurrently; the next stage waits for the shared context it needs. This separates parallel work from the dependencies between rounds.

A provider abstraction connects OpenAI, Anthropic and Google models. Agent roles and prompts define their perspectives, while a debate store publishes updates through server-sent events. The OpenAI Apps SDK widget presents messages, votes and the selected answer in the chat interface.

The coordinator parses structured votes and handles ties. Those mechanics make the selection process inspectable, but agreement between agents is not proof that an answer is correct. Additional model calls also introduce latency and cost.

What shipped

An MIT-licensed MCP server and interactive widget demonstrating a complete debate, from initial proposals to vote results and a winning answer. Screenshots and the implementation are available in the public repository.

The result is an exploration of agent orchestration and transparent interfaces, with a concrete workflow that makes each stage of the discussion inspectable.

Architecture

MCP tool

Receives the question and starts a tracked debate.

Coordinator

Runs proposal, critique, refinement and voting in order.

Model providers

Executes the five agent roles through a common LLM interface.

Live widget

Receives event updates and displays contributions, votes and the result.

TypeScript · mcp-use · OpenAI Apps SDK · React · Server-sent events

Explore the project

Explore the project through its demo and available resources.

• Source code & demonstration
• Debate coordinator